TSMC 3nm Overload
TSMC’s 3nm capacity is described as “overloaded” as orders from Nvidia, Apple and hyperscalers outstrip available wafer starts — customers are now battling for allocation and non‑AI products face rationing. That squeeze is forcing companies to prioritize buffer inventories and accelerate yield optimization planning for fabs and integration sites like Apple Fremont. (technode.com)
Reports from multiple industry outlets say TSMC has been prioritizing wafer starts for its longest‑standing, highest‑volume partners — notably Apple and Nvidia — leaving mid‑tier and newer customers facing delayed allocations. (techinasia.com) Market analysts report N3 lead times and package lead times stretching beyond 50 weeks and estimate TSMC’s 3nm capacity is effectively booked for roughly the next 18–24 months. (siliconanalysts.com) Advanced packaging (CoWoS) and HBM memory supply are identified as the immediate chokepoints: CoWoS is described as oversubscribed through at least 2026 and HBM3/HBM3e allocations are reported fully committed for enterprise accelerators. (fusionww.com) Analyst models peg a 3nm wafer‑start cost near ~$20,000 — about 1.8× the cost base of mature 7nm wafers — a move that is already widening BOMs for AI accelerators and premium mobile SoCs. (siliconanalysts.com) TSMC is expanding capacity geographically: company planning and local reporting indicate use of its second Kumamoto fab for 3nm production with reported investment figures around $17 billion, alongside accelerated plans to scale advanced‑packaging throughput. (money.usnews.com) Apple has accelerated U.S. manufacturing commitments under its American Manufacturing Program and added four domestic partners — Bosch, Cirrus Logic, TDK and Qnity Electronics — in a March 26, 2026 update tied to its broader multi‑hundred‑billion U.S. investment pledge. (cnbc.com) Supply‑chain analyses warn major AI customers are front‑loading wafer purchases and explicitly advancing yield‑engineering roadmaps, with modelling showing a single GW of next‑gen GPU deployments could consume a very large share of N3 capacity and drive customers to accelerate yield work. (isaiahresearch.com)
